VeriSpoof-Plus-MLX-8bit

Beyond real or fake: one unified audio LLM for speech deepfake forensics.

VeriSpoof-Plus in MLX for Apple silicon Macs, with an 8-bit language model and audio encoder. Same prompts, scores and API as the PyTorch release: detection, generator identification, source verification, recording condition and grounded reasoning.

	Size	In-the-Wild EER	Same generator (ASVspoof 5)	Same acoustic model (SpoofCeleb)
VeriSpoof-Plus (PyTorch, bf16)	6.0 GB	1.00%	19.50%	18.67%
VeriSpoof-Plus-MLX-8bit (this model)	2.9 GB	1.00%	20.00%	18.50%

600 recordings or pairs per set from VeriSpoof-Bench.

Quick start

hf download AustinXiao/VeriSpoof-Plus-MLX-8bit --include "code/*" --local-dir verispoof-mlx
pip install ./verispoof-mlx/code

from verispoof_mlx import VeriSpoofMLX

model = VeriSpoofMLX.from_pretrained("AustinXiao/VeriSpoof-Plus-MLX-8bit")

model.detect("sample.wav")              # {'label': 'spoof', 'score': ...}, score > 0 means spoof
r = model.explain("sample.wav")         # grounded reasoning + per-token synthetic-speech probability
print(r["text"])
model.same_source("a.wav", "b.wav", level="generator")
model.condition("sample.wav", attr="codec")
model.generator("sample.wav", ["XTTS-v2", "Bark", "VITS"])

License

CC BY-NC 4.0 (non-commercial use), as VeriSpoof-Plus. The base models keep their Apache-2.0 licences (LICENSE-Apache-2.0.txt, NOTICE).

Citation

@misc{verispoof2026,
  title         = {{VeriSpoof}: Beyond Real or Fake with One Audio {LLM}},
  author        = {Xiao, Yang and Dang, Ting},
  year          = {2026},
  eprint        = {2611.04634},
  archivePrefix = {arXiv},
  primaryClass  = {eess.AS},
  url           = {https://arxiv.org/abs/2611.04634}
}